- Who it helps
- Researchers studying how extra reasoning affects image composition and prompt following.
- Availability
- October 30 paper and public project page. The inspected repository still lists inference, training and evaluation code as pending; its expected November 2 inference date is not proof of release.
Reason in the model’s working representation
Training gradually replaces token-based reasoning with internal states, then uses reinforcement learning to refine them. The authors report improved image quality and prompt following, including a seven-percentage-point GenEval gain over their BAGEL comparison.
LARC studies image reasoning through internal model states ↗
A research result with an availability gap
The project links examples and model resources, while the inspected code repository contains only a README. Reproducing the full method remains unresolved in this check. Gains on these benchmarks do not establish a universal image-generation improvement.
LARC studies image reasoning through internal model states ↗LARC release checklist ↗LARC project explanation ↗
What you can do
- Check the repository release checklist before treating this as a runnable method. Open reference ↗
Where the evidence stops
No model or evaluation executed. Independent replication, complete release contents, licenses and deployment costs remain unresolved.
Sources read. Primary documentation read November 2, 2026 UTC. No software installed, model inference run or full workflow executed.
Sources & further reading
Paper · paper
LARC studies image reasoning through internal model states ↗
Identifies the subject, documented capability or specific verification limit.
Byungwoo Jeon and coauthors · page checked 2026-11-02
GitHub · implementation
LARC release checklist ↗
Identifies the subject, documented capability or specific verification limit.
Byungwoo Jeon and coauthors · page checked 2026-11-02
More sources reviewed (1)
Web · official
LARC project explanation ↗
Identifies the subject, documented capability or specific verification limit.
Byungwoo Jeon and coauthors · page checked 2026-11-02
Coverage note: Independent attention and end-to-end execution remain unverified.
The record over time (1)
Dated source checks and published explanations. An observation date can differ from the original release date.
- · Added to Current
LARC replaces part of a multimodal model’s written reasoning with continuous internal states before generating an image.